OpenAI has announced GPT-5.5-Cyber, a cybersecurity-specialized version of its AI, and says it significantly outperforms Anthropic's rival model on a cybersecurity benchmark. According to a Bing News item citing OpenAI, the company claims GPT-5.5-Cyber surpassed Anthropic's "Mythos 5," sharpening an already intense rivalry between the two labs over security performance.
The launch is part of a broader push. According to The Decoder, OpenAI is expanding its "Daybreak" cybersecurity initiative with three pieces: an updated Codex Security plugin, the full GPT-5.5-Cyber model, and a partner network that the outlet reports includes more than 25 security firms along with several governments.
Perhaps the most notable shift, as The Decoder describes it, is in what the technology is meant to do. Rather than concentrating on finding software vulnerabilities, the focus moves toward patching them automatically — closing security holes without waiting for a human to write and apply the fix.
The sources here are announcements and coverage of OpenAI's own claims, so the benchmark comparison reflects OpenAI's framing rather than independent testing. No specific scores, testing conditions, or details about Anthropic's response were provided in the source items.
Why it matters: cybersecurity is one of the clearest real-world tests of how far AI can be trusted to act on its own, and a tool that moves from spotting flaws to fixing them automatically — backed by security firms and governments — signals how quickly that responsibility is being handed to machines.